This Content Downloaded From 141.211.4.224 On Sun, 03 Jan 2021 11:18:46 UTC

WHY DO SURVEY RESPONDENTS DISCLOSE MORE WHEN COMPUTERS ASK THE
QUESTIONS?
Author(s): LAURA H. LIND, MICHAEL F. SCHOBER, FREDERICK G. CONRAD and HEIDI
REICHERT
Source: The Public Opinion Quarterly , Winter 2013, Vol. 77, No. 4 (Winter 2013), pp.
888-935
Published by: Oxford University Press on behalf of the American Association for
Public Opinion Research
Stable URL: https://www.jstor.org/stable/24546177
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
and Oxford University Press are collaborating with JSTOR to digitize, preserve and extend
access to The Public Opinion Quarterly
This content downloaded from

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
All use subject to https://about.jstor.org/terms
Public Opinion Quarterly, Vol. 77, No. 4, Winter 2013, pp. 888-935
WHY DO SURVEY RESPONDENTS DISCLOSE MORE

WHEN COMPUTERS ASK THE QUESTIONS?
LAURA H. LIND
MICHAEL F. SCHOBER*
FREDERICK G. CONRAD
HEIDI REICHERT
Abstract Self-administration of surveys has been shown to inc

respondents' reporting of sensitive information, and audio compu
assisted self-interviewing (ACASI) has become the self-administra
method of choice for many social surveys. The study reported he
laboratory experiment with 235 respondents, examines why A
seems to promote disclosure. It compares responses in a voice-only
administered) interface with responses to a face-to-face (FTF) hum
interviewer and to two automated interviewing systems that prese
animated virtual interviewers with more and less facial movement. All
four modes involved the same human interviewer's voice, and the vir
tual interviewers' facial motion was captured from the same human
interviewer who carried out the FTF interviews. For the ten questions
for which FTF-ACASI mode differences (generally, more disclosure in
ACASI than FTF) were observed, we compared response patterns for
the virtual interviewer conditions. Disclosure for most questions was
Laura H. Lind is an independent researcher, San Francisco, CA, USA. Michael F. Schober is a
professor of psychology at the New School for Social Research, New York, NY, USA. Frederick
G. Conrad is a research professor in the Survey Research Center and director of the Michigan
Program in Survey Methodology at the University of Michigan, Ann Arbor, MI, USA, and research
professor and director of the Joint Program in Survey Methodology at the University of Maryland,
College Park, MD, USA. Heidi Reichert is a statistician at the University of Michigan Medical
School, Ann Arbor, MI, USA. The authors thank Shelley Feuer, Don Harrison, Rebecca Rosen,
Dave Vanette, and Huiying Yan for assistance in data collection, analyses, and formatting. This
work was supported by National Science Foundation grants [SES-0551294 and SES-1025645
(Methodology, Measurement, and Statistics program) to M.F.S. and F.G.C.]; by the Charles
Cannell Fund in Survey Methodology at the University of Michigan through a grant to L.H.L.;
and by New School for Social Research support to L.H.L. and M.F.S. * Address correspondence
to Michael Schober, New School for Social Research, Department of Psychology, 80 Fifth Avenue
Room 710, New York, NY 10011, USA; email: schober@newschool.edu.
doi: 10.1093/poq/nft038 Advance Access publication November 21, 2013

DThe Author 2013. ! Published by Oxford University Press on behalf of the American Association for Public Opinion
Kesearcn. /\n ngnis reservea. ror permissions, piease e-man: journais.permissionsvyoup.com

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Disclosure to Computers vs. Survey Interviewers 889
greater under ACASI than in any of the other modes, even though th
two virtual interview modes involved computerized self-administratio
This suggests that the locus of FTF-ACASI effects is particularly tie
to the absence of facial representation in ACASI. Additional evidence
suggests that respondents' affective experience (e.g., comfort) durin
the interview may mediate these mode effects.
Survey respondents tend to report more socially undesirable behavior wh

questioned by computers than by interviewers. Population estimates
cocaine and marijuana use, for example, are greater when respondents answ
anonymously via audio computer-assisted self-interviewing (ACASI)—list
ing to recorded questions on headphones and entering answers on a lapto
so not even the interviewer can hear—than when they answer the same qu
tions asked by a person (e.g., Tourangeau and Smith 1996; Turner et al. 199
Reports of the average number of opposite-sex sexual partners in the past yea
in the past five years, and since the age of eighteen are similarly affected by
mode of administration: women report more partners to a computer than t
interviewer, and men report fewer (Tourangeau and Smith 1996; Tourange
et al. 1997). In general, the evidence suggests that the increased reporting
socially undesirable behavior under computerized self-administration refle
greater response accuracy; for example, respondents' reports of their poor
demic performance (low grades, dropped classes, academic probation) mo
closely matched official records in a web survey than in a telephone interv
(Kreuter, Presser, and Tourangeau 2008).
These patterns can emerge whether questions are spoken (prerecord
voice) or textual (e.g., Tourangeau and Smith 1996; Couper, Tourang
and Marvin 2009). They can emerge whether the spoken questions are pr
sented in ACASI or telephone touchtone response systems, also know
T-ACASI or interactive voice response (IVR) (e.g., Tourangeau, Coup
and Steiger 2003; Steiger and Conroy 2008); they can emerge textua
in web surveys and on the laptop during what is otherwise a face-to-fac
interview (text-CASI). Evidence like this has led to widespread adoption o
audio- and text-CASI for collecting data on almost every kind of potenti
sensitive topic.
What is it about automated self-administration that leads to greater rep
ing of socially undesirable behaviors? Or, conversely, what is it about hum
interviewers that leads respondents to be less forthcoming? Any one of
many differences between human interviewers and automated interview
systems could in principle produce FTF-ACASI disclosure differences. Th
purpose of the study reported here is to begin to identify which features
combinations of features are actually responsible.
Understanding the causes of FTF-ACASI mode differences is parti
larly important because they are not always observed. Within a sing

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
890 Lind et al.
survey, some qu
ference while o
observed reliably
text-CASI than with face-to-face interviews but no mode difference for life
time drug use. And a mode difference can be found for a question in one
study but not for a similar question in a different study. At present there is
no systematic way to predict when a mode difference is likely to emerge—
for which questions, for which populations, or for which implementations
of a mode. Nor is there a systematic way to decide when to deploy comput
erized self-administration for those respondents, questions, or surveys for
which it will be most effective. Understanding the causes of FTF-ACASI
disclosure differences is an important step toward anticipating when they
are likely to occur.
Potential Origins of FTF-ACASI Disclosure Differences

The general explanation for FTF-ACASI disclosure differences is that the
presence of an interviewer motivates respondents to present themselves
in a positive light and to avoid reporting potentially harmful information.
Privately reporting socially undesirable behavior to a computer seems
to be less threatening than reporting that same information to a human
interviewer, and it may reduce worries about subsequent disclosure of
compromising information to third parties such as family members, law
enforcement, or tax offices (Aquilino 1994; Schaeffer 2000; Tourangeau
and Yan 2007).
Note that this line of thinking is at odds with the pervasive notion that peo
ple are more likely to disclose embarrassing information to a nonjudgmental
person with whom they have rapport; this assumption is at the heart of psy
chotherapeutic practice, forensic interviewing techniques, and Alfred Kinsey's
in-depth interviewing methods. From this perspective, one would imagine that
a human interviewer should elicit more reports of sensitive behaviors than a
computer (to the extent that survey interviewers are perceived in the same way
as Kinsey sex researchers, police interviewers, or psychotherapists). Although
this is not the predominant finding in comparisons of computer and human
administration, a few comparisons of different human interviewing modes
(FTF versus telephone) suggest that greater interviewer presence can some
times lead to more disclosure. For example, in studies of smoking (Gundersen,
Delnevo, and Momperousse 2007), weight (Béland and St-Pierre 2008), and
driving after drinking alcohol (Béland and St-Pierre 2008), respondents were
more likely to report socially undesirable behavior or facts to an interviewer
face-to-face than on the telephone, where the presence of the interviewer is
less evident. It is unclear whether the greater reporting of this sensitive infor
mation to face-to-face interviewers results from greater rapport with the

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
respondent, the respondent's awareness that the interviewer might be able

verify the answer through physical observation (smelling cigarettes, see
body mass, etc.), both, or something else. But it does suggest that interview
administration might sometimes lead to more disclosure than computeriz
self-administration: a conventional computer surely cannot detect the smel
cigarettes or the respondent's BMI.1
A fruitful path to more fully understanding why ACASI and other modes
self-administration promote disclosure is to decompose the interviewing si
ation into potentially relevant elements and systematically manipulate the
This follows the long tradition in human-computer interaction of decompos
"humanness" into features that are added incrementally to interfaces to ma
them more human (e.g., Sproull et al. 1996; Yee, Bailenson, and Rickertse
2007).2 It also further tests Tourangeau, Rips, and Rasinski's (2000, 306-7
proposal that "efforts to humanize the interfaces of computer-assisted metho
of self-administration may, by creating a virtual human presence, increase
level of misreporting about sensitive topics."
Here we investigate the possibility that the differences between intervie
ers and ACASI systems could contribute independently, or in combination,
respondents' willingness to disclose sensitive information. As table 1 outlin
we see a number of important differences between interviewers and AC
systems, and there are no doubt others. Any one or some combination of th
differences could be responsible for mode effects in the reporting of sensit
behaviors to ACASI systems and humans. For example, the primary reason
the difference may be that ACASI systems do not have evaluative capabilit
and that respondents are averse to disclosing when there is the potential f
judgment or disapproval—independent of whether the interviewer is em
ied (FTF or virtual), has dialogue capability, or has agency.3 Or perhaps t
primary reason is that ACASI systems are not visually or verbally respons
to the respondent's answers beyond advancing to the next question when
answer has been entered, and it is the lack of interviewer responsivity t
promotes willingness to disclose.4 Or perhaps embodiment is the key: perh
1. If sensory technologies (e.g., Gutierrez-Osuna 2004; Person, D'Mello, and Olney 2

Nakamoto 2013) are deployed on a larger scale, respondents' expectations of computers'
bilities could change, potentially reducing the likelihood or direction of ACASI-FTF resp
differences in these domains.
2. For additional background, see Preece, Rogers, and Sharp (2011) on the controversies a
anthropomorphism in interaction design, and Cassell (2000) on the complexities of embody
human characteristics (e.g., improved dialogue capability, speech output and recognition, em
ied human facial attributes) in interfaces.
3. See, for example, Ellingsen and Johannesson (2008) for evidence from behavioral econo
that the potential for verbal evaluative feedback can affect altruistic behavior.
4. In everyday conversation, speakers are well known to be affected by how their conversati
partners attend to their utterances and provide interactive feedback (for a review, see Schober
Brennan [2003]).

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
892 Lind et al.
_ _ bo c
y In C o CO
C
0)
00
Vl
gs-3
^ «U
>.
wen
ed :p 3
I-O
"23"£
S ^
^•
•^3-0 O £ t Ji O £ 03 C -ê O
<D
*t=i
D
22
•<!>*-<>
-h "O _o
Ch
tfl
*^3 </j
®
S c
«\
U
^ S «-h aj aj : o ^ w S O S .
.2 to ° ^'a&> o <*> Ti C .2 T! 9 g
s:§S.>.§^
«
° 1 •s s - B Bg
° & r I
si| §>
2 « iu ir» n w «C C
2 '«
ll I! til S II | rill
r: W 3 >. ^ (N I a< g H 3 c 2 °"
« u o x: « ~ -.3 o gfc 8 «i -5 u
^ a o . S S 9 & ^ . §5^5
° -3 & o £ £ g fc
■g ? g
•nuKEo— | •:§
u^*® u a, £ -g
I? ■s -a <2
S e «! . > -
1&8.b|^J£S| I -a a 1 ! 111!
S S cm »S 3 | >■» §=2 " § » 3 2
•2"3.'S.oo ^ •§! ^5!t«^S
gg us?
Si"" |u §o fi
£ u'o
s
■gi E5 Eo£ & e-° £ c o >> EEc-B-S
0,-73 >,.2 ^ S ■» 5 ^ ^■S « -o
S -a £ 55 £ £ S g 'g
c/5w5<Zltic^'^0^<y5^ c/j O 4) <, <zj <zj ^ ^
_H-^h-l^»-Hr^rd(UM(D I—\ Cu W rj KH H W b -'
g g a <
<1 .S ^ o «s
interviewingsytemsofthefut re). w .S ^ .S <^*So H ^ o o .h ^ y s ^ « g | y ylla
itsown,deci not adminster hequstionaire, tc.
M 'o _ & S "

£ ^ -8" S, "is § " S -
o_^S ~1*jm
">; . a u .c
% . B -a J <S "§ § « f £ g -S *
■fill 13 | .if ssBI 11 I g £
lis ill ll 3 !| i! ?! 5 f I
•S illi r|| |i pi 11 Ml !!
-o •§ 2 iS o ^ <d ^ -u .c 13 .S o o %
w ^ o J c -3 c o- ?? ^ « S 611 > *-» (D g
« g '£ .23 =3 S p E „ '3 aS « . 2
UQ.S T3 J3 § 6 o C K £ § "2 *CU-g
(fl m „ -p
Cw) 
„ W 2,—«
—g1/2
9 2
.52 8Sg-a
•£»
•—) *s f|
^ 2 -8 g-S
* -3 :=T t: £ £ c S O g rf J8 73 'C oS
.&;«§ js§ ss 5^.22 tl
W)
53 .2
5 £ <UC
5 o £ "5 -g •" £ #U3 >,S . s ^ JS a
^ ^ cn 2 x: H «4-hC
0 X o g « ■= VI C"S3 «;°ta?'-f=s"^':S&2c
odDhS >^^Sl « iflCS u e 11 „ .2 u o
^ D t_ T3 ^ J O c O "S o O x; fc 73 cdo
o^SoRb! o-o^c^c^'jS&^-Sccû
"I £-§§ ? 8 ! SSSs^S 1| 3 s§3
tegfe|fe=do J3«« »|2iS»c 6i s SoS
1 3 I ^ J I % -g (?.§ ! 11 £ g 'g | g ! !
c^cu
m-SwC
1111 ill in 11 §521 ii i i§l
Interviwscan,poteialy,produce(siptdorunscipted)ACSIsytemdontgeralypoducerbso neutralpobesrvnclarif tonased.clarif ton(alhuginprclethycould). cos
cognitveadfctivesa whilespakng dwhileandMr[208],165-;Foucalt[209];vonderPitn listengoaswerndterighmntohecmputr.eal[201]fordiscu nofpteialdspyb^>3"
tSoa d auditorycuesoftheirdmographic arteisc—theirag,nterviw'sdemographics. ecJ3^>Ccc
virtualO aBytheirvypesnc,itervwsdiplaybothvisualndACSIsytemonlyhaveuditorycesfthreco<i th►Lio'tJcsM.-
nic ty, gender, etc. them(validornt)aboutherspondet'scognitveandcoul;sePicard[198]). afectiv stae,thepysical onditonadcirumstances, and emographic arcteis c tha mightberlvant o interpreting response . cue Interviwcandisplyorhidee-eySoM choset,andcotiungobealtcnomaterwh coletshanwers;itcanofrminteos danythig respond toes.l norex cisd retionabutwha od;itcano ctn
vncofthajudgment.ACSIsytemcanotdisplayorcnealjudgmnt. muC
U X .1 3 -S -I X
o | '"% & S | 3 c
S & G u& > & ">
Table1.DiferncsbetwnHumanIterviw sandTypicalACSIystem Dimenso HumaniterviwerACSI DialogueIntrviwpoducesriptdueancsireltminwaysACSIpechisreodanechqustiondelvr capbiltyhawilnevrbxactlyhesmforqS
alespndt.iprecslythamewyvrtime sadinter. EmbodientIrviwe'sbodyanfce movandexprshACSIsytemhavnobdilyrfacdisply(eCasl m a.PercptualInerviwcanobservth spondet'hysicalACSIsytemcanobservomakeinfrcsabout capbiltyharceistandplys,andcmkeinfrcsomtherspondt(huginprclemoS
advncesytm EvalutieInrviwehast bilytojudge(positvlyrnegativly)heT ACSIsytemcanotjudge. capbiltysenitvebhaviorstharespondetsrpotinheintrvew.
pjo< AgencyItrviewcantofherwnacod,mvingwhesACSIystemonladisterhuveyqstionad

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
respondents' willingness to disclose sensitive information is reduced by t

mere presence of the interviewer's face.5 There is certainly a long-stand
practice of promoting disclosure by concealing visible evidence of on
interlocutor, e.g., in the confessional booth or on the psychoanalytic cou
although obviously those situations differ from survey interviews.
The study reported here was designed to start distinguishing between th
possibilities. In the study, we compared responses to questions asked by a
human interviewer to responses to questions asked by three comparable co
puter interfaces that were derived from attributes of the same interviewer. A
interfaces used speech by the same interviewer, with recorded speech in t
computer-based conditions. The ACASI interface presented questions usin
only recorded speech. In the two other computer conditions, an animated f
presented spoken questions; facial motion was captured from recordings of
FTF human interviewer and projected onto the animated face, thus creatin
"virtual interviewer." In the High Animation (HA) virtual interviewer in
face, multiple channels of motion were projected onto the animated face
the Low Animation (LA) version, only the lips moved while the rest of the
was motionless. The four interfaces differed on more than just a single feat
(see below); the logic was that these comparisons would allow us to rule o
or in major hypotheses about what causes ACASI-FTF disclosure differen
so that future investigation might make finer, feature-by-feature distinction
If disclosure is greater to all the computer interfaces than to the human in
viewer, this would suggest that the primary locus of ACASI mode effects
in respondents' belief that computers cannot understand (they lack dialog
perceptual capability) or negatively evaluate (they lack evaluative capabili
or intentionally disclose information,6 and their belief that humans are q
tatively different from computers. We will call this the Human Interview
hypothesis. If disclosure to any interface with a face (human or virtual in
viewers) differs from the audio-only interface, this would suggest that em
iment—here, the mere presence of a face—reduces disclosure, regardless
whether it is biological or digital. We will call this the Facial Representa
hypothesis. If disclosure is affected by the degree of facial animation—say
disclosure were equally curtailed by a high-animation interface and a hum
interviewer—this would suggest that a primary locus of ACASI-FTF mod
5. Faces seem to play a crucial role in the activation of interpersonal schémas, with special st
in neural representation (e.g., Kanwisher, McDermott, and Chun 1997), and in newborns' pre
ences (e.g., Johnson et al. 1991) as well as their cognitive capacities (e.g., Johnson and Mo
1991). Eyes may be particularly important: people can behave more prosocially when they
watched by even static images of eyes, littering less (Ernest-Jones, Nettle, and Bateson 2011)
voluntarily paying more for coffee with an "honesty box" (Bateson, Nettle, and Roberts 200
6. Note that respondents' beliefs about the confidentiality of their answers could go either w
respondents could be more candid with a computer because they believe it will not disclose s
sitive information, or they could be more candid with a human because they trust the hum
confidentiality assurance more than a computer's (as Aquilino [1994] has argued).

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
894 Lind et al.
differences is hu
that they are hum
Our study comp
that allowed us t
across the differ
capture technolog
is that the same human interviewer's voice and facial motion could be used
as source material for all the computer interfaces—the same interviewer who
administered the FTF interviews. This allowed for fair comparisons across the
conditions.
The strategy in the current study was to administer a set of questions that
had either led to mode differences (human interviewer versus self-administra
tion) in large samples or that plausibly could lead to such mode differences.
The analytic approach was then to compare patterns of response across all
four modes (FTF, HA, LA, and ACASI) for only those questions that had
demonstrated a FTF versus an ACASI difference. Thus, we could examine
whether the mode effects resulted from the interviewer's being human rather
than a machine (Human Interviewer hypothesis), the interviewer's motion or
responsivity (Animation hypothesis), or the fact that there was a moving face
versus no face at all (Facial Representation hypothesis).
Various patterns of results consistent with these hypotheses are possible.
The clearest pattern would be a single "dividing line" for all questions that
produce a mode effect. For example, if all three automated systems pro
duce the same response distributions and they differ from those produced
by the human interviewer, this would make a clear case that the locus of
FTF-ACASI mode effects is the presence or absence of a human. Equally
straightforward would be a pattern showing a single dividing line elsewhere
(e.g., between modes with faces versus ACASI). An alternative possibility
is a monotonie increase in socially desirable answers from ACASI to FTF,
suggesting that there is an independent contribution of each additional set of
features of humanness to respondents' reluctance to disclose. It is also pos
sible that the dividing line is different for different survey items, suggesting
that the cause of FTF-ACASI mode differences may vary across questions.
Finally, it is possible that more than one hypothesis is supported for any one
7. The impact of animation on users' experience and cognition has been investigated in interface
domains other than surveys, including education (e.g., papers in Lowe and Schnotz [2007]) and
comprehension of graphs (e.g., Tversky, Morrison, and Betrancourt 2002). A particularly effec
tive use of animation coordinates gesture and speech of embodied animated agents (Cassell 2000;
Cassell et al. 2000).
8. One could certainly study the locus of ACASI mode effects through mode comparisons other
than those used in this study. For example, one could compare responses from FTF interviews,
video-mediated interviews (e.g., via Skype or FaceTime), and video recordings of human inter
viewers. Or one could compare live human audio interviews (telephone), advanced automated
speech dialogue systems, and simple ACASI.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
question; this would suggest that multiple causes for FTF-ACASI mode dif
ferences can jointly affect a response.
Experiment
DESIGN AND PROCEDURE
Respondents were randomly assigned to one of four experimental cond

tions. They participated in a forty-two-question interview conducted in the
same environment, a university laboratory in New York City, by either 1) a
human interviewer (FTF); 2) a High Animation (HA) virtual interviewer
3) a Low Animation (LA) virtual interviewer; or 4) an Audio Computer
Assisted Self-Interviewing (ACASI) system. (Our implementation of
ACASI differed from typical administration in that there was no section
of the interview conducted by an FTF interviewer, nor was an interviewer
in the room during the self-administered sessions.) In all conditions, a
human experimenter (not the interviewer) brought respondents into the lab
and oriented them to the interview before it started. This experimente
provided the same confidentiality assurances to all respondents to reduce
the possibility that respondents' trust in data confidentiality would differ
across modes.
After the interview, the experimenter returned and accompanied respond
ents to a different room, where they completed a text-based questionnaire in
a web browser (not used for computer-administered interviews) that asked
respondents to rate how they experienced the interview (e.g., their comfort,
their enjoyment, and how natural they found the interaction). In addition,
respondents answered questions about their experience with computers in
general.
In the FTF condition, the respondent and interviewer sat facing each other.
The interviewer read aloud questions presented on the screen of a laptop com
puter. Respondents answered by speaking, and the interviewer entered the
responses into the laptop. In the three computer-based conditions (HA, LA,
and ACASI), each respondent sat alone in the same laboratory room used in
the FTF condition, facing a desktop computer and wearing headphones. In the
HA and LA conditions, the virtual interviewer was displayed on the computer
screen above clickable (radio button) response options or a data entry field
into which respondents could type a number, depending on the question (see
figure 1). There was also a "Repeat Question" button to re-present the question
and a "Next" button to advance to the next question. In the ACASI condition,
there was no visual display of the virtual interviewer; response options and
the "Repeat Question" and "Next" buttons were displayed just as in the HA
and LA conditions while respondents heard the same audiorecorded questions
presented in the HA and LA interfaces.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
896 Lind et al.
O Excellent
O Very good
O Good
O Fair
O Poor
O Don't know
O Not sure
Repeat Question
Figure 1. Virtual Interviewer User Interface (with more facial movement

in HA than LA condition).
INTERVIEWING INTERFACES
The HA and LA virtual interviewers were created using motion-capture

ware (proFace Video) for extracting motion from a moving human face.9
table 2 lists in detail, the HA and LA virtual interviewers differed in seve
ways so as to make the HA interviewer more like a human interviewer;
9. The process involved three main steps. First, blue and green foam dots were affixed t
cific places on the human interviewer's face. Second, a head-mounted video camera worn by
interviewer and aimed at her face recorded her face (including the dots) while asking the su
questions chosen for the study. Third, the movement of the dots was extracted from the rec
video via the motion-capture software and applied to a computer-animated head so that the h
interviewer's prerecorded facial movements controlled the virtual interviewer's face.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
"H S o g << o g I
C/3
<
g >h Z >h ££££££ £ £ Z^g
H
a!
flj O w 5.1)^
C3
c3 (j ^ Z ^ z ^ z z z z ^ z ^ ^
H
£ aJ
O
J
"O
*p C/5 a> tt! C/3 (*> C/2
<= £ £ £ £ £ £ z £
Interviewing agents <L>
W) &4
^ <zj <zj c/5 m <| i/i M m w «3n c« y>
<L>
£1 ££
Face-toftL,
acr8
eHighanimtonLwanimtonAC SI
c
o
S 0) 3
0)
^ cr
 c c
Vh
.,
W
i/i
fl
a.
C/3
1-1
8--3
■<-»
•-* co
a | a §
c .a - -o s a
—
E O
5,O3u
ea
«J
S 3
CT*
J c ffl u r o t-1
^ 2 1 ^ c *
-..22 u « d, g
ni S o i^ 0 S 2P I*
.
= O-§ A)
S-^SS
5 ^
d..x»r5
_.„>« -a 3 « " £ a
■Oj3 3 E "g £ ,2 "a 2 >.£ a £
ills o o & -s *8 s «H
? S S £
5 ^K<S«~ g
° as *co Ou_,"
g „ ™ £5ja
J2 ?s
II I g -pi £ 8 ! J |2§S
I I! £ 1 i 11 s i M!
1 ? B IS
f j§§^S SI)
1 Sg.-g
»■ 2 G ? 1| S||a !5
? 2855§:a®iii
I 1aI «a1"3a§"2o73i 0)
IU Sj 'S
gointaskyou mequstionabutpersonalfi ces.")Y es Y 1 I 'S a « "S lmotin)whlerspondetisanwerigqustion®Yes No/A c/5"I -a ^s ^■LAinterfacdisplayedsmouthmventasHA,butnoherfacilmovent. -a J a 2
Table2.FeaturesofInterviewngA ents
^3 2,
q.
Audio features
y J3
3 0J] 2 0 t2
blinks,fac(ii
3 MU'OoTik'SS^S c/3 33
SpechLiveRcordeRcordeRcorde QuestionrepatvilbeYs Yes Questionrpeatdwihntro(i.e,"Ltmerpathequstion.YesNo Transitonsbetwnsurveyquestionpics(.e,"NowI'm
fli O
Visual features Facilrepsntaio fnterviw Yes YesNo Fade-inofvde clipsN/AYesNo /A MouthmvesinycwithspecYse N/A HeadmoventYs No/A EyeblinkgYes No/A FacilexprsionYes No/A Contiualevidnceofatenio (.e,hadmovemnt,ey ContiualgzetrspondetNoYes N/A Interaction RespondtaswerbyspeakingYesNo RespondtableorfusetoanwerqustionYes Yes Modefnorespn romptSokenp TextualTextual AcknowledgmntofrespndtacionsSpkeno Nneo

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
898 Lind et al.
online appendix C
supplementary mat
applied to the com
the captured motio
eye blinking, and n
presented entirely
RESPONDENTS
Two hundred thirty-five respondents from the New York City area we
recruited via an online advertisement (on Craigslist New York City) and
print and online ads (in the Village Voice), each of which offered $20 for
ticipating in a survey. The sample consisted of 132 women and 103 men,
a mean age of thirty-one years. The respondents were randomly assign
one of four experimental conditions, with fifty-nine assigned to the FTF
dition, sixty to the HA condition, fifty-seven to the LA condition, and f
nine to the ACASI condition. There were no significant differences amon
groups in age, race, level of education, or experience with computers.10
SURVEY QUESTIONS
Forty-two questions were selected from three well-known and consequential

ongoing U.S. social surveys: the Behavioral Risk Factor Surveillance System
(BRFSS), the Survey of Consumer Finances (SCF), and the General Social
Survey (GSS). The questions concerned seven different topics: health and
fitness, sexual health and behaviors, alcohol use, leisure activities, political
activism, personal finances, and altruistic behaviors. Of the forty-two ques
tions, twelve were chosen because they had produced mode effects (inter
viewer administration versus self-administration or FTF versus phone) in
large-scale surveys, and twenty-one were chosen either because socially desir
able responding had been observed in other studies or because we judged some
responses to be more socially desirable than others, and so it was conceiv
able that mode effects could appear in our smaller lab sample. Nine additional
("neutral") questions for which no mode effects were expected were included
to make the survey feel cohesive. (See appendix A for the full list of questions
and the reasons for each question's inclusion).
For the twelve questions that had produced mode effects in large samples,
we derived predictions about the reporting of socially undesirable behaviors.
10. Experience with computers was measured in several ways: self-reported time spent reading
and sending email per week; time spent on the web for other tasks; time spent in chat rooms,
newsgroups, or on bulletin boards; use of Facebook and other social media; experience with ava
tars in online settings; self-reported Internet expertise; and a combined digital literacy measure
(Hargittai 2005) based on questions asked in the Internet Modules of the General Social Survey
in 2000 and 2003.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
On ten of the questions, respondents in the published studies reported mo

socially undesirable or embarrassing behaviors in the more anonymous mo
and so we would expect respondents to report more of those behaviors in
ACASI condition than FTF. For two questions (smoking and weight),
previously demonstrated mode effects (Béland and St-Pierre 2008) were i
the opposite direction than has been typically observed, with more report
of sensitive information FTF than in telephone interviews. If we assume t
telephone interviews afford respondents more privacy and distance from
interviewer than FTF, we should see a similar reversal in the patterns of m
differences here, with more reports of smoking and being overweight to F
interviewers than to ACASI. For the twenty-one questions that we deem
potentially likely to produce mode differences in socially desirable respo
ing, we would expect more candid (less socially desirable) responding
ACASI than in FTF. For the nine "neutral" questions, we would expect
differences in responding FTF versus ACASI.
STATISTICAL METHODS
We used ANOVA to test mode differences in the twelve questions r

ing open numerical (continuous) responses; logistic regression for the e
questions that required yes/no responses; ordinal logistic regression fo
nineteen questions requiring selection from an ordered response scale (
Every day, Several times a week, Several times a month, Rarely, or Never)
multinomial logistic regression for the three questions requiring a choice
an unordered response scale. For six numerical questions (male sex part
female sex partners, sex partners in last year, sex partners since age eightee
new credit card charges, and drinking regularly), the response distributio
skewed to the right; for these questions, we log-transformed the data and e
ined not only the full distribution of responses, but also the response pat
for the top quartile of responses. Focusing on respondents whose reports
largest makes sense on theoretical grounds: a respondent who never drin
has never had a sex partner, for example, cannot report less drinking or
sex partners, while a respondent who drinks frequently or has had many
partners has more "room" to underreport the total with a human interv
or report a higher, more accurate total in ACASI. And presumably those
have more to report experience more social pressure to misreport, althou
course norms may vary in different subpopulations. So mode effects that
not emerge in a distribution with many zero values may be visible amon
subgroup with the most to report.
11. This is a composite created by summing reported male and female partners since age eig

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
900 Lind et al.
Results
QUESTIONS THAT PRODUCED FTF VERSUS ACASI MODE EFFECTS
Recall that our analytic strategy was to compare patterns of response across
all four modes only for those questions whose answers differed FTF versus
ACASI. For the nine questions for which we did not expect mode differences,
there were no reliable differences between these modes. For the thirty-three
questions for which mode effects seemed plausible, answers for ten of them
differed significantly or marginally between FTF and ACASI, including one
(sex partners since age eighteen) for which the effect was found for the top
quartile of responses but not the full distribution; these are the questions on
which we will focus. (It is not apparent from the content of the questions why
these items produced the effect and the others did not.12) For nine of these
ten questions, the FTF versus ACASI difference was in the expected direc
tion based on previously published mode effects or findings on socially desir
able responding. (For eight of these nine, we expected more socially desirable
responses FTF than ACASI; and for one, smoking, we expected the oppo
site.) For the remaining question (drinking regularly), the difference was in the
unexpected direction: more regular drinking was reported in FTF than ACASI.
The leftmost columns of table 3 (FTF versus ACASI) summarize these effects.
RESPONSE DISTRIBUTIONS
Table 4 presents the summary measure (mean, median, or percentage)

each of these ten questions that, in our judgment, is most revealing abo
patterns of responding. For example, the question about watching prim
television offers the following ordered response options: "every day," "
times a week," "several times a month," "rarely," and "never." For thi
tion, table 4 presents the percentage of respondents who reported wa
primetime television several times a week or more. (See appendix B fo
exhaustive response distributions.)
Where did responses to the virtual interviewers fall relative to FTF
ACASI? As table 4 and appendix B show, for all ten questions response
both virtual interviewers fall in between the FTF and ACASI responses.
TESTING THE THREE HYPOTHESES
We carried out a series of comparisons between our experimental condition

(see table 3) to test our hypotheses. To test the Human Interviewer hypothesi
12. Note that our observing FTF-ACASI differences for ten of the thirty-three sensitive questio
in the survey is consistent with previously reported studies, where mode differences are most often
found for a subset of sensitive questions in a survey—sometimes for only one or two.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
c cd
.2 « "S
CA cd
*o3 c3 On O T3 .y
o o -g ' '
o S ^ (fl te
o o
Continued
<, £> <D Oh
c/J &l
> <D O
c« 04 P*
0> ^ p
O C3 ^
NO on o o ,T,
* '§ ^ffl
C/3s NO <N m CO ro w
• • • • l-N ry,
CO
CJ> O CJ5 O
C3
§ -2 II
§ S3
NO <w>
8 -o -a
Q 5
1! S--H pa" o o
o
p<
>>
■ 43 x in h- ^
High animation CQ Tt <N r-H (SJ pq

^ o c^> o ^
« s
£ £
D. (1) ^
-!
c3 c3 O O 2 • ~
'§ §•
o o
£ '? -22
O K <*)
0 ^ i>
$&%
§1£
1 I" NO On <N On r -
m n IT) (N rr\ W
Humanvs.computervs.lowanimtonFacesv.ACSI (HumanIterviwer(Animaton(FacilReprsntaion
«8
cj> o cj> d ^
CO cd
<
<
-1
« a
OO T-H ^ O
2 o3 3 "S
C0
§ o o
> £ "
BOds dBO s d
* . . it —«.
m «n t-h no . .
BPartileaBPrtiale BPartileaBPr«s
tiale t CO NO CO rQ w
?(0.35) (0.29) (0.2S?
5)(0.29) (0.36)(0.29)(0.26)(0.3) (0.41)(0.35)(0.28)(0.32)
odd
FacetofaP-<
cer*
vs.ACSIhypotesi)hypotesi)hypotesi)
3
O
o<
ffi
>
TVHours-0.61# 0-.H
52#01-.70 2-.30 4
HaveCrditars0.89*2430.51740.2170.63*18
<D
Q*
"O I !
LogistcSmoke1.06*2.90 481.620 *1.8209*2.5 reg sion(0.4)(0.3)(0.27)(0.34)
o Z o
Table3.ModeCmparisonfrtheTnQuestionShwingaSnifcatFT-ACSIEfect ModeltypSurveyqstion(SE)squared(SE)squared(SE)squared(SE)squared ANOVSexPartnsl8±-0.75*<8-0.36 -0.48# 6-0.* 9 ModeltypSurveyqstion(SE)ratio(SE)ratio(SE)ratio(SE)ratio

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
902 Lind et al.
High animation
Humanvs.computervs.lowanimtonFacesv.ACSI (HumanIterviwer(Animaton(FacilReprsentaion
BPartileaBPrtiale BPartileaBPrtiale
Facetofacevs.ACSIhypothesi)hypothesi)hypothesi) (0.34)(0.28)(0.24)(0.27) (0.3 )(0.27)(0.23)(0.27) (0.34)(0.28)(0.24)(0.28) (0.3 )(0.28)(0.24)(0.16) (0.3 )(0.27)(0.24)(0.27)
Savings0.64#1.90 3 1.480321.8047#1.60
TVPrimet -0.73* 48-0.5# 64-0.25 78-0.53# 9 Seatrnge0.72* 05.47#160.21 30.48#162
GivngHomels0.6*1820.3190.7190.34*N/A°
Note.—Mdlsarep ntdbyhepofrsne:ANOVmodelsfrnumical(ontus)repon,lgistcre onmdelsfry/noesp ,and "Ordinalogstcmdelwrfitusngheloitnkxcepithasewrthpaleinsumptionfaled:Fcsv.ACSIestimaforGvingHmels 1Forupeqartileof spndet(umberofsxpartnes>=20),coefintsrepo dnthelogsca. #p s 0.10; *p £.05; * p<.01
Table 3. Continued ModeltypSurveyqstion(SE)squared(SE)squared(SE)squared(SE)squared OrdinalReguarlyDink0.62#180.43150.91 0.415 logistc(0.3)(0.27)(0.23)(0.26) reg sion0Newspaer0.94*2.5604*1.90 48*1.620 *1.86 ordinalgstcre ionmdelsfro despncal.Prmetsia(B)regivnwthsadreosinparthes(SE).fctizesarlopentd. questionbasedonprbitlnk;coefi ntsforlgitnkmodelsrpotednlog-dscale.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
<D
3£
3 g
&c ^ ^ co
£ ;T ON .
- ^ ^ (N —< rH 00 CM
oo © m »o-h ô oo m <N
.2
"3^_ V, <N 00 ^ ^ ^ <*>
p «
Audio computer
c
.2
c3 00 ^ rf* Tt r-H ON r-.' i-H O
£ °i in © in © oo
w w ^ ^ co m (N (N co
in r- co mo oncsvo as
£ ^ ^ i" <N (N <N
O
H-l
c
.2
as tF ^ P p co" prop ^ ©
£ °i ^ \6 in oo o oo in o in
WW WW www w w
_ — ino co o h >n r- o\
-gj rM' ^ ^ <N<N co —
•*1
CN oo (S co oo in vq o
© ON h OO O in K
in -rf co <n co co —i
(N
,. _ © on cNr-oo
03 CN rf CO (N CN
Face-tofaceHighanmtionLwanimtonasitedslf-intervw
S
= * £ *! _ * ! 1 *
oo "Tj S ~ ^ Z "S ^
£ ^ £ 5 \ ^ w
3
>. uo ■§ £>
a .2 > S
o % £ U.c§ j>
g —i *
«c u 73 u „
j= p £
^ _ * 3
=a«*g|
"O *3 ^ a
u t <, o- « ^ ^ ^ § o « ~ ^
^ !» S « fe' S ■* S u"
> I I S. I S3 g I -g I i §§
.1 B11-S* flliSS*?!
tU-s^§ |&s|lS^-§s
sl^..ss »s
andSE)2.(04)2.8(04)2.5(08)2.(07)
a p p -5 Sofe-'aa ^ «j"e-S * .2 53
birthday(median,N)41.5(6)45(1)47(16)80(14) (jV,%)30(5.8)27(45.0)23(4.)15(2.4) awekorm e(N,%)2(37.)0(5.)29(50.)36(1.0) awek(N,%)21(35.6)17(30.4)19(3.0)12( .0) personatlesonceawk(N,%)10(7.)9150(.8)17 "Top quartile.
Table4.Sum arizedPaternsofRespondi g HourswatchingTVperd£
ay(mean NZ Havecrditars(N,%)46(78.0)4(6.7)43(5.)3(59.) Smoked10 cigaret sine tireb!
umberofsexpartners ince18th Im life Readingewsparevyda(N,%)£
29(4. 38)20(35.1 (8.6) WatchingprimetimeTVsevraltimes mSavergulayechmont(N,%)17(28.)17(28.3)1(2.)Q
8136 Drink3tmesawkorme(N,%)18(30.5)1(2.0)168 5(.) GivngseO
at oastrangeratleastonce O
Givngfo dormoneytoahomel s

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
904 Lind et al.
we compared re
To test the Anim
and ACASI respo
compared ACAS
face was represe
As shown in tab
nificantly in on
ported in two m
of these question
hypothesis supp
ence of a living
FTF and HA interviewers—that drives the FTF versus ACASI effect.
The Animation hypothesis was supported in two questions (smoking and
newspaper reading) and marginally supported in one more (sex partners since
age eighteen), although it was not the only hypothesis supported for these
questions. The responses for HA and LA virtual interviewers (table 4) seem
more similar on these questions than one might expect if the amount of anima
tion is the locus of the FTF-ACASI effects. To examine this, we carried out
two additional comparisons between 1) responses to the two virtual interview
ers (High and Low Animation); and 2) responses to the two virtual interview
ers and FTF. There were no reliable differences for any of the questions in
either comparison. So, the support for the Animation hypothesis in these data
may be a statistical artifact rather than an explanation: if amount of animation
is responsible for the FTF-ACASI difference for these questions, then surely
responses to HA and LA should differ and responses to the automated systems
should differ from FTF, but they do not.
The Facial Representation hypothesis, in contrast, is supported for eight of
the ten questions that had shown an FTF versus ACASI difference; it is sup
ported significantly in five of the ten questions, and marginally supported in
three more. The Facial Representation hypothesis is the only hypothesis sup
ported for three of these eight questions (Credit cards, Savings, and Giving to
the homeless), which suggests that the presence of a face is indeed the locus of
the FTF-ACASI effect for these questions. For the other five questions, other
hypotheses are also supported: Animation for three and Human Interviewer
for three (both Animation and Human Interviewer hypotheses are supported
for the sex partners question).
13. See online appendix D for all comparisons—FTF versus ACASI and the hypothesis tests,
as well as two additional comparisons (ACASI versus HA and LA, and HA versus LA)—for all
forty-two questions in the survey. As the appendix shows, there were twelve marginally signifi
cant (p < .10) and four significant (p < 0.05) effects out of the 242 additional comparisons beyond
those presented in table 4. This is no more than would be expected by chance for this many com
parisons and suggests that our analytic strategy of testing our hypotheses on the ten questions for
which FTF-ACASI mode differences were observed is appropriate.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Underscoring the notion that all modes with faces tend to group together
are the additional analyses indicating no differences in responses to the two
virtual interviewers and FTF. Because of this, the Animation hypothesis seems
less viable as an explanation in the cases where the Facial Representation
hypothesis is also supported. The locus of the FTF-ACASI mode difference
for these questions (three for which only the Facial Representation hypoth
esis was significantly supported and two for which both Animation and Facial
Representation hypotheses were supported) thus seems to be the presence ver
sus the absence of a face (live or virtual).
The one question for which the FTF versus ACASI difference was not in
the expected direction (drinking regularly) did not yield support for any of
the three hypotheses. There were no other reliable differences between the
conditions in any other comparisons, although for the top quartile of responses
there were several effects we could not interpret; we do not report them here
because there were no FTF-ACASI differences among these responses. We do
not have a direct explanation for either the reversal or the lack of support for
any of the hypotheses, but note that the frequency of regular drinking in this
sample was quite low. It may be that this question was therefore not as sensi
tive for these respondents as has been seen in previously reported findings, and
that the difference does not reflect socially desirable responding. Given this,
the reversal of the FTF versus ACASI mode effect may not allow real testing
of our hypotheses for this question in this sample.
Overall, the patterns demonstrated different "dividing lines" (discontinui
ties in disclosure between different groupings of modes) across the ten ques
tions, with more than one hypothesis supported for some questions. This
clearly argues against an account of FTF-ACASI mode differences based
only on whether a live human interviewer asks the questions and records the
answers (the Human Interviewer hypothesis). The hypothesis receiving the
most support (in eight of the ten questions, uniquely in three) was the Facial
Representation hypothesis, suggesting that even a clearly nonhuman facial
representation with limited motion can reduce disclosure as much as an actual
human interviewer relative to a voice alone.
More generally, the pattern is consistent with our suggestion that human
interviewing comprises a bundle of features which, when deployed indepen
dently in survey interviewing systems, can independently affect disclosure.
What is less clear is why certain combinations of features affected answers for
the particular questions that they did—e.g., support for the Human Interviewer
hypothesis for TV hours watched, but support for the Facial Representation
hypothesis for having a credit card.
We don't see anything in the content of these questions that would lead to
one pattern of results rather than another, although surely the content matters.
In fact, we suspect that there may not be a simple mapping between domain
of questioning and the kinds of mode effects we are documenting here. Can
the patterns instead be explained by how respondents experienced the different

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
906 Lind et al.
interviewing mo
the virtual inter
positive (e.g., "I
to the more enth
good teeth smil
I enjoyed"). If re
rassed or uncomf
ing agent (along
and Yan [2007]), o
with faces, for e
interview than t
examined the po
POSTINTERVIEW RATINGS
We had complete postinterview questionnaires for 213 of the 235 respon

ents: fifty-seven FTF, fifty-eight HA, fifty-one LA, and forty-seven AC
Prior analyses that had shown FTF-ACASI differences in the full sample w
repeated on this subset and resulted in similar patterns of mode differen
We interpreted this as indicating sufficient similarity between the subsam
and full sample to justify further analyses.
To the extent that there were differences in ratings across modes, FT
and ACASI respondents rated their overall experience of the interview m
positively than either HA or LA respondents. Respondents rated their in
action as more natural (on a five-point scale) in both FTF and ACASI
HA and LA (for FTF compared to HA and LA, F[2,203] = 10.44, p < 0.
for ACASI compared to HA and LA, F[2,203] = 3.74, p = 0.03). They
rated the ACASI interviewer as more like a human (as opposed to m
like a computer) on a three-point scale than the HA and LA intervie
(F[2,148] = 5.31, p = 0.01); the HA and LA interviewers were rated equ
human, (F[l,148] = 0.63, n.s.).
Respondents rated the FTF interviewer more positively than the
puter-based conditions. Respondents reported that their comfort le
increased over the course of the interview more (on a three-point scale)
FTF (/• [ 1,211 ] = 5.25, p - 0.02); that they enjoyed the FTF interview m
(on a five-point scale) than the computer-based interviews (F\ 1,212] = 9
p < 0.01); and that they found the FTF interview less frustrating (on a f
point scale) than the computer-based interviews (F[l,210] = 10.56,p = 0.00
These patterns of ratings do not seem consistent with the disclosure findin
The fact that respondents' ratings of the interviewers' naturalness in FTF
ACASI did not differ is inconsistent with our observed FTF-ACASI disclosure
difference and with any patterns of disclosure that we observed across the four
modes. The ratings of comfort, enjoyment, and frustration are in one sense
consistent with the Human Interviewer hypothesis in that they differentiate

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
between the human interviewer and the computer-based modes, but they s
at odds with the findings on disclosure; one might expect respondents w
disclose less to report greater discomfort, less enjoyment, and more frustrati
in FTF than in the computer-based modes. Perhaps respondents who disclo
less feel happier about the interview.
To further clarify the relationship between the disclosure patterns an
respondents' reported feelings about the interview, we looked more closely
differences among respondents within the experimental groups. We dich
mized respondents into those who reported being very or extremely comf
able at the start of the interaction and those who did not, as well as into th
who reported enjoying the interview somewhat or thoroughly and those w
did not, and those who rated their interaction as more natural and less na
ral. (Of course, these variables do not measure precisely the same underlyi
construct, but they all can be argued to reflect comfort or a sense of hum
connection.) We then tested whether the FTF-ACASI difference in disclosu
(for the ten questions that had shown a difference) interacted with any of th
three dichotomized affect measures.
Figure 2 shows the pattern that we observed for the question about giving
a seat to a stranger. Respondents who reported having been more comfortable
at the start of the interview showed a typical FTF-ACASI mode effect on dis
closure: respondents were more likely to report that they had not given a seat
to a stranger in the past year in ACASI than FTF. Respondents who reported
not having been comfortable showed no effect of mode, interaction of comfort
(yes-no) and mode (FTF-ACASI), Z = -2.90, p - .004. Of course, we do not
know the direction of causation: on the one hand, respondents who felt com
fortable with the FTF interviewer could as a result have been more concerned
about how the interviewer perceived them and so disclosed less than respond
ents who felt uncomfortable with the interviewer (e.g., less connected, less rap
port) and thus did not mind disclosing the socially undesirable behavior. On the
other hand (reversing the causality), respondents who had little to disclose (or
felt unembarrassed about not giving a seat to a stranger) may have felt more
positive about the interaction, while respondents with more to disclose may
have felt worse. The same causal ambiguity holds for ACASI: respondents who
felt comfortable with the ACASI system may have felt uninhibited and thus
disclosed more, or respondents who had much to disclose may have felt more
positive about their disclosure when made to an automated system.
Either way, this finding suggests that how respondents experience the inter
view is related to how much they disclose, and that ACASI-FTF mode effects
may be driven by a subset of the respondents. We observed the same pattern
for three more of the ten survey questions (two significantly and one margin
ally) on at least one dimension of reported experience (naturalness, comfort,
humanness). Clearly, these data are more suggestive than definitive, but the
patterns are consistent enough to justify further investigating an experience
based account of the FTF-ACASI mode effect.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
908 Lind et al.
How often have you offered your seat on a bus or subway,

or in a public place to a stranger who was standing?
0.20
>■ 0.15
0.10
■ Face-to-Face
■ ACASI
2 0.05
0.00
No Yes
Comfortable at start of interview?
Figure 2. Marginal Probability of Giving the Least Socially Desirable

Response ("Not at all in the past year") for Respondents Who Reported
Being Comfortable vs. Not Comfortable at the Start of the Interview.
Discussion and Implications

It has long been known that interviewer administration of survey question
(compared to self-administration) can suppress disclosure and increase socially
desirable answers, presumably because respondents wish to avoid embarrass
ing interviewer reactions and other potentially harmful consequences, such a
disclosure to family members or law enforcement. The current study furthe
explores what it is about the presence of a human interviewer that is respon
ble for survey mode effects of this type, both to better understand the origin
of these effects and to potentially predict when survey modes, whether current
or future, are likely to affect answers. The proposal is that different aspects
of human interviewers—embodiment, perceptual capability, dialogue capabi
ity, voice, etc.—can independently affect respondents' willingness to disclose
or to misreport, and thus that FTF-ACASI mode effects do not result simply
from the presence or absence of a human interviewer. Instead, the particular
combination of these attributes in a survey mode leads to a particular level o
disclosure.
Our findings largely support this proposal. Specifically, they demonstrate

that embodiment in the form of a face even with limited motion can increase
socially desirable responding (consistent with Sproull et al. [1996]) just as
much whether the face is biological or digital—the Facial Representation
hypothesis—and even in a relatively small sample. Even though the presence

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
of a human can drive the mode effects (as it did uniquely for our survey q
tion about TV hours watched), disclosure can be inhibited by interfaces t
have only some features of a human interviewer. Beyond this, the findi
rule out certain possible explanations for the causes of FTF-ACASI m
effects. For example, support for the Facial Representation hypothesis can
result from an unnaturally fixed gaze by the interviewers: HA and LA int
viewers gazed fixedly at respondents, but the FTF interviewer's gaze var
because she needed to look at her laptop; nonetheless, responses in all thr
modes grouped together for eight questions. Similarly, support for the Fa
Representation hypothesis cannot be attributed to how respondents registe
their answers. In all three automated modes (HA, LA, and ACASI), respon
ents entered their answers textually and with the mouse while FTF resp
ents spoke their answers. Yet, the automated modes with faces grouped w
FTF and not ACASI.
Our results extend the domains of questioning for which FTF-ACASI
mode effects have been observed. For example, prior studies on mode effects
in answering questions about smoking have observed more reported smoking
FTF than on the telephone (Gunderson, Delnevo, and Momperousse 2007)
or more reported smoking in paper-and-pencil self-administration than FTF
(Brittingham, Tourangeau, and Kay 1998); here, we see an FTF-ACASI
mode effect with greater reports of smoking not only FTF but in the self
administered modes with a face (HA and LA). This is curious if the explana
tion for the mode effect is that respondents feel they can't lie about smoking
to a live interviewer because the interviewer can smell smoke and see tobacco
stains; our virtual interviewers had no such perceptual capability. Perhaps
embodiment—digital or human—evokes a sense that the interlocutor has per
ceptual capabilities that of course on closer inspection would be dismissed.
Alternatively (and equally curious), in the domain of smoking, respondents
may have experienced a level of rapport with a digital, moving face equiva
lent to their rapport with a live interviewer, leading to similar increases in
disclosure.
Of course, we do not know which reports by respondents are true; the fact
that less socially desirable answers in other domains tend to be more accurate
(Kreuter, Presser, and Tourangeau 2008) increases the plausibility of interpret
ing our results as we do, but there is obviously no guarantee. Nonetheless, the
findings are compelling given that they emerged in a relatively small conveni
ence sample, i.e., with limited statistical power, and the patterns of data are
generally as one would expect given prior evidence (with the exception of
drinking regularly). The experimental design also reduces the plausibility of
alternative explanations for the pattern of results; it is unlikely that the results
are due to differences in respondents' motivations to participate (respondents
were randomly assigned to conditions, and they all received the same mon
etary incentive), different response rates across the conditions (respondents
did not know which condition they would be assigned to when they made the

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
910 Lind et al.
decision to partic
confidentiality o
before the surve
across all four co
The findings rep
picture about th
interface that re
web interface does not seem to affect disclosure of sensitive information
(Tourangeau, Couper, and Steiger 2003). Recorded moving faces, however,
do seem to reduce disclosure, at least when accompanied by a voice, com
pared to disclosure through a textual interface; for two additional examples,
see Sproull et al. (1996), who found higher social desirability scores and
reports of greater altruism to talking heads than to text, and Fuchs (2009),
who found reduced disclosure of sexually transmitted diseases and certain
sexual behaviors, relative to text, when a recorded video interviewer's gen
der differed from the respondent's, but in some cases greater disclosure to a
same-gender recorded video interviewer. This starts to suggest that there is
something special about moving, speaking faces—but they don't have to be
human, live (as opposed to recorded), of particularly high fidelity, or particu
larly reactive to respondents' behaviors. From the other end, it suggests that
what is special about ACASI is that the recorded speech is disembodied—it
lacks a moving face.
The proposal that interviewing modes are collections of features raises the
possibility of designing automated interviewing systems that strategically
tailor particular combinations of features to maximize respondents' willing
ness to disclose sensitive information. One might implement the same set
of features for all respondents, different features for different subgroups of
respondents, different features for individual respondents, and even differ
ent features for different topic domains or questions within an interview.
Potential benefits include increasing (some) respondents' engagement, effort,
and completion rates by including particular features (like a moving face)
for some respondents (e.g., the kind of respondent who reported enthusiasm
for our virtual interviewers) or for some questions (e.g., the smoking ques
tion or "neutral" questions). Potential drawbacks include reduced disclosure,
disengagement, or increased breakoffs for other respondents (e.g., the kind
who found the virtual interviewer creepy) or other questions (e.g., altruistic
behaviors).
To implement this, ideally one would know, for each respondent in a sample
and for every question, several key attributes. First is the respondent's general
propensity to disclose. Some respondents may be predisposed to disclose in
any mode, or about particular topics in any mode, and others may avoid dis
closure in general or about a particular topic in any mode. Respondents in the
middle on this disclosure propensity continuum should be most likely to be
affected by features of an interviewing mode.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
A second attribute useful to know would be the likelihood for a respon

ent that the question is sensitive. Respondents who have more to hide m
feel particularly embarrassed or threatened with a more humanlike inte
viewing interface, and more comfortable when human features are minim
Respondents who feel they have nothing to hide are less likely to be affe
by interview mode. Some topics are likely to be sensitive for everyone (
income questions in the United States and Europe), while questions on ot
topics may be sensitive only for subgroups (e.g., heroin use is unlikely to
sensitive for nonusers) (see Schaeffer 2000).
Another potentially useful attribute would be respondents' affective r
tions to an interface. Respondents may vary in how much rapport or emb
rassment they feel interacting with a human interviewer or with a human
interface.
Knowing how these attributes are distributed across the sample for a
ticular question could allow one to predict whether mode effects are likely,
thus whether an adapted interface might be effective. In our view, mode eff
should depend on the proportion of respondents who are affected by mode
the size of the mode difference for those who are affected, just as nonresp
error depends on both response rate and the size of the difference in wha
being measured between respondents and nonrespondents (e.g., Groves e
2009). From a practical perspective, some of this information is inferable f
published evidence about general trends in the population and, potential
from data about individual respondents collected in prior waves of a pan
study or earlier in a single interview.
The current study does not compare all possible humanizing features t
could affect disclosure. Our virtual interviewers involved one head paire
with one voice with movement of one human actor; whether the effects
would be the same with a different human actor or different interface char
acteristics needs to be verified. Although the effects emerged across sev
eral topic domains, it is also unknown when and how they extend to other
domains. And the fact that multiple features differed across the experimental
conditions does not allow conclusions about the relative contributions to
disclosure of those different features. Nonetheless, the findings demonstrate
that the well-known differences in responses to human interviewers and
self-administered computer systems do not result simply from the fact the
computers aren't human; rather, individual or bundled features of human
ness in a user interface—especially a facial representation—can still reduce
disclosure.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
912 Lind et al.
Continued
St)
— — 'C 32 5/3 "G

c —
3 £ £ o ^ c o
O OD D O W)
a, -r CL, 
S, -g
>» o
® E
to
PublishedNomode
.2 r« = c $ w
sizdef ctsDepnd t3C

SurveyofmdeHypothec/d Mode f ct *3 3 >» QQ K
o CO » on cn w CQ
lifts |
C3 £" C 03 £
2 ^ fee- T3 U 3 . £ ^ 6
« S3 c ckSm^SSoO.h'u'3
GO
e J5 op •£ g O £ j c
"1 o « 13 £ si S3 -5
& TO <D .5 3-> <L> to
>-j= - g Q a 3 «- n P «r> c a «
tl 5 •£ s> xi -* ~ 'a1 £ o « c s ^ ^ « «
& « 3 % -3 8 E :, 1 >> 5 # p 1 S 5 s
s-c-oS.^ UXÔS'-Ofe ^'So
OSOÊC uUo-S.SOÔu^jsS
>-, oj O J; r2 2 rzl 4J • — c/j 1— 1-1 i/~~i 3
JV «J M O 2 --3 CX -O q ^
- •>- & 3.H"SS^!SS^S
O O V C OS«c3p^^^«^D,b"0
X3 -3
Appendix A. Questionrgefctmodefctxpe dvariblepuishdMO lXOrde yourhealt isexcl nt,veryRiskFactor espone god,god,fairpo>>
oesDirctonMdelots Wouldyousayth inge ralBehaviors>
JT, -> O O C Q O *-« TI "* ^ "S .C! ^ *-■
r,don'tSurveilance atgories know,oryou'renotsure?System (Geoo
n ralHealth) (BRFS ) genralBRFS XOrde physicalexm,notan<ai:o.t;«aZ?,a>>o
Aroutinech kupisa^ examrespone foraspecif njury,ilnes,categories orconditon.Howlonghas itbensinceyoulastvisted adoctororotherhealthcare providerforaroutinecheckup? Nevr,5ormoreyearsago, withinthepast5years,within thepast2years,withinthepast year,withinthepast6months, ornotsure?(SeDoctor)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Continued
PublishedNomode
SurveyofmdeHypothesizdef ctsDepnd tMode f ct
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots HowdyuescribyourBRFSXOrde Béland CAPI/Lower weight?Veryunderwight,respon St-Pier 208CATIrepotsfobe slightyunderwight,abouctegoris tyinCATI the right weight, slightly over weight, or very overweight? (Weight) OnhowmanyofthepasBRFSXNumericBéland CAPI/Signfcatly 7daysi ouexrcisoSt-Pier 208CATIlower pots partic enphysical tviy(Tourangeuofxercis n foratles20minuteshatl.197hadCAPI madeyousweatndbreath verysimlar hard,suchabsketbal,soc-findgwitha cer,unig,swim nglaps, ightlydifernt fastbicylng,fastdncig,question their orsimlareobicatvies?boguspieln (DaysExercis)expriment)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
914 Lindetal.
Continued
Delnevo&FT mode Momperous e 2007
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepuishdMoesDirctonMdelots Think gaboutn riton,hwBRFS XNumeric many to al servings of ruit and/or veg tables di you eat yesterday? A serving would equal one medium ap le, a handfulofbroc oli,oracupof car ots. (FruitVegServings) Haveyousmkedatls10BRFSXYes/noBéladnCAPI/Higher cigaretsinyourenti lfe?St-Pier 208;CATIrepotsf (Smoke)Gunders ,moking HaveyouevrbentsedforGenralSocialXYes/no HIV?Dontcount est youSrvey(GS) mayhavehad spartofablo d donation. (HIVTest)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Continued
and female
for women in SAQ
Smith19 6higher eports
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots HowmanysexpartnshaveGSXNumericTouange FT/HigherLo youhadintelas12months?eal.197;SAQrepotsfr ansformed (SexPartnesYar)Touangeu&meniFT; Duringthepas12months,GSXUnorde Fournder haveyoursexpartnesbnrespo categoris: exclusiveymale,xcusivelycatgoriesn ,male femal,orbthmalendoly,fema femal?(MFSexPartns)oly,bthmale Abouthwoftendi youhaveGS XOrde sexduringthelast12months?reponse Nota l,once rtwice,once atgories a month, 2-3 times a month, we kly, 2-3 times per we k, 4 or more times per we k? (HowOftenSex)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
916 Lindetal.
Continued
FTF
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepuishdMoesDirctonMdelots NowthinkgaboutheimGSXNumericTouange FT/HigherLo sinceyour18thbir day,etl.197;SAQfemaltrnsfomed howmanyfemalprtnesTourangeu&rpotsin haveyouhadsexwith?Sm 196SAQthani (FemalPrtnes18)FT NowthinkgaboutheimGSXNumericTouange FT/HigherLo sinceyour18thbirday,howetal.197;SAQfemaltrnsfomed many lepartneshaveyouT rangeu&repotsin hadsexwith?(MalePrtn sl8)Smith196SAQthani NowI'mgoin toread listofGS XUnorde termstha peol someti sreponse usetodescribethmselv :categories het rosexual or straight; homosexual, gay, or lesbian; andbisexual.Whichtermbest describes how you think of yourself? (Sexual denti y)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots Adrinkofalch isonecaBRFSXNumericAqulno&SAQ/Higher orbtleofbr,oneglasLoSciut190CATIrepotsf ofwine,o canorbtleofdrinkg winecolr,necoktail,orSAQmode oneshotofliquor.Duringthe past30days,howmanydays perwek(orpermonth)di youhaveatleastonedrink ofanyalcohlicbevrage? (DaysDrinkPast30Days) ConsidergaltypesofBRFSXNumericAqulno&SAQ/Higher alcohibevrags,howLSciuto190CATIrepotsf manytiesduringthepasdrinkg 30daysi ouhavemorSAQmode than5drinksononeoc asion? (MoreThan5 Drinks)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
918 Lind et al.
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepuishdMoesDirctonMdelots ThinkgbackoverthlasBRFSXOrde Aquilno&SAQ/Higher 12months,abouthwregulayrespon LSciuto190CATIrepotsf diyoudrinkalcohibevr-categorisdnkign ages?Nvrin12months,1-3SAQmode timesin12months,4-7times in 12 months, 8-1 times in 12months,1-3timesamonth, once or twice a we k, 3-4 times per we k, 5 times a we k or more? (RegularlyDrink) Again,syouthinkbacoverGSXNumericAqulno&SAQ/HigherLo thelas12months, wmanyLoSciut190CATIrepotsfransomed drinkswouldy haveon drikng typicaldywhenyoudrank?SAQmode (TopicalyDrink) Duringthepast12months,GS XYes/no have you read novels, hort stories,poems,orplays,other thanthoserequiredbyworkor scho l? (ReadNovels)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Continued
watching in CATI
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots How ftendoyureadtheGS XOrde newspaer—veryda, fewrspone timesawek,onceawek,les categoris than once a we k, or nev r? (Newspaper) Ontheavrgeday,bouthwGSXNumericBdges20 Mail/Lower manyhoursdyoupersnalyCATIrepotsf watchelvison?(TVHours)telvison I'lnowaskboutsmedif r-GSXOrde entkindsoftelvison hows.reponse Wouldyoutel m how ftencategoris you watch primetime drama or situation comedy programs? Would you say ev ry day, sev raltimesawe k,sev ral times a month, ra ely, or nev r? (TVPrimetime)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
920 Lind et al.
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots How ftendoyuwatchworldGS XOrde ornatio lnewsprogams?reponse Everyda ,sevraltimesactegories we k, sev ral times a month, ra ely, or nev r? (TVNews) Andhow ftendoyuwatchGS XOrde progams hown publicrespone telvison?Everyda,sevralctegories timesawe k,sev raltimes a month, ra ely, or nev r? (TVPublic) Whicofte lowingSurveyofXUnrde Colapsedto staemnsbetdscribeConsumer sponethrunoder youcrentmployentFiaces(SCF)categoriscategoris: stau?Employedfrwages, mployed, self-mpoyed,utofwrkvoluntariy formethanoeyar,outnemployd, ofwrkolesthanoeivluntariy year,homeakr,studen, mployed retired, unable to work? (EmploymentStatus)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Continued
PublishedNomode
Ap endix A. Continued Questionrg efctmodefctxpe dvariblepuishdMoesDirctonMdelots NowIhavesomequstion SCFXYes/no aboutcreditcardsandcharge cards.Doyouhaveanycredit cardsorchargecards?Please do not include debit cards. (HaveCreditCards) AreanyofthecardsyouhaveSCFXYes/no anytypeofVisa,MasterCard, Discover,orAmericanExpres cards you can pay of over time?(Donotincluderegular American Expres charge cardsthatmustbepaid nful.) (CredictCardsPayOverTime) Howmanydo uhave?PlaseSCFXNumeric do not count duplicate cards for the same ac ount or any busines or company ac ounts. (HowManyCreditCards)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
922 Lind et al.
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots Onyourlastbi,roughlyowSCFXNumericLog muchwert newchargstnformed madetohseacounts?valuerpoted (NewCharges dholers)focardholes After h lastpymentswer SCFXNumeric made on thes ac ounts, roughly what was the bal nce stil owedonthes ac ounts? (Balance) Think gonlyaboutVisa,SCFXOrde MasterCad,Discover, spone AmericanExpres cardsyoucategoris canpayof overtime,andstore cards, do you almost always, sometimes, or hardly ev r payof theto albal nceowed ontheac ounteachmonth? (CreditCardsPayof Total)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots NowI'dliketoasyuomeSCFXOrde Colapsedto questionabutyoratiudesrponsethr oder aboutsaving.Peoplhavectgoriescatgories: difernt asonfrsaving,do'tsave, evnthoug eymanotbesavoc savingltheim.Whicofsnaly,sve thefolwingstaemntscomergulay close t odescribngyoursaving habits? Don't save—us al y spend more than income; don't save—us aly spend about as muchasincome;savewhatev r isleftoverat he ndofthe month—no regular plan; save incomeofonefamilymember, spendtheother;spendregular income,saveoc asionalother income; save regularly by put ing money aside ach month? (Savings)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
924 Lind et al.
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepuishdMoesDirctonMdelots Overth pastyear,wouldSCFXOrde yousaytha yourspendigrespone exc de yourincome,tha ctegories it was bout he same as your income, or that you spent les than your income? (Spending) Somepoles mtofl wSCFXOrde what'sgoin o ingovern- espone mentadpublicaf irsmotcaegories of the time, whether ther 's an el ction going on or not. Othersaren't hatinter sted. Would you say ou fol ow what's going on in government and public af airs most of the time, some of the time, only now and then, or hardly at l? (Fol owPolitics)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots In204,youmayrem berGS XYes/no thatKer yranforPresidenton the Democratic ticket ag inst Bush for the Republicans. Did you vote in that el ction? (Vote2004) Whatboulcae tions—doGSXOrde youalwysvoteinthose,dorespone yousmeti smisone,docategories you ra ely vote, or do you nev r vote? (VoteLocal) Inthepast hre o furGS XYes/no years, have you at ende any politcal me tings or alies? (PoliticalRal y) Inthepast hre o furyeas,GS XYes/no haveyoucontributedmoneyto a politcal party or candi ate ortoanyotherpolitcalcause? (PoliticsMoney)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
926 Lind et al.
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots Duringthepast12months,GS XOrde how ftenhaveyoudnatedrspone blod?Nota linthepastc egoris year, once a month, two r thre times a year, once in the past year, or don't know? (DonateBlo d) Duringthepast12months,GS XOrde how ftenhaveyougivenrspone fod rmoneytoahmels categoris person? Not at al in the past year,onceinthepastyear,two or thre times a year, once a month,onceawe k,morethan onceawe k,ordon'tknow? (GivingHomeles )

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Continued
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots Duringthepas12months,GSXOrde Acre tal.19 Mail/Hgher howftenhavyoudnersponeCATIrepotsf volunterwokfrachity?caegorisvlunter Nota linthepasyer,wokin onceithpasyer,twoCATI or thre times a year, once a month,onceawe k,morethan onceawe k,ordon'tknow? (Volunte rWork) Duringthepast12months,GS XOrde how ftenhaveyougivenrspone moneytoachrity?Nota ctegoris al in the past year, once in the past year, two r thre times a year, once a month, once a we k, more than onceawe k,ordon'tknow? (MoneyCharity)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
928 Lind et al.
PublishedNomode
Ap endix A. Continued Questionrgefctmodefctxpe dvariblepushdMoesDirctonMdelots Duringthepast12months,GS XOrde how ftenhaveyou ferd espone yourseatonabusorinactegoris public place to a stranger who was tanding?Nota l inthe pastyear,onceinthepastyear, two rthre timesayear,once amonth,onceawe k,more thanonceawe k,ordon't know? (SeatStranger)

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
r- co r~
_ <*1 . © ON ©
55 S 2 © Q S Continued
52 oc 00 in ^ in (N
< rxi CO <N Tt
00 \c Tf VO
O Tt OtN
£ © i£ K WW W W
>n ^ n t cn o
< ^ ^ - <N m
o , , , , , , , , , ,
o Tt t— CO © ©
_ °i O vd <ri iri ©
=5 £ O ^ VOC2- S ^
<f oo o o t-~ o
g rsi <N (N m
Tj- O O 00 <N
(SI
00 (N O 0\
J5S- i2 —' WW WW
cs \o cn o o\
« rn (N
a "5
52 r>
Q C*3 CV.
! 1 * &
CV. ^ O, "<■ ^
-5
s S
.s -e
Q s ^2^-k
s: ^
O^S? ^ ^ ^ s j?
■§11 I I Si =§"* I £
•> ""•>» .. "î Q C«o 'Kt s
J3 b-S S a" t
~ Cs 2?-c
t_o?-3
^ vS 53 §L
£?"§ -51 «■ ^ Si « He
N59 7 N16 4
8i g a ^ ® || § ^ 3 "8 *
§ « § -3 c | I I |il I II| B ll|g#
3 2s© ^22 t « a
ApendixB.CompletRsponeDistrbuonsfrTenQustionShwingFT-ACSIModeDifrnces
'S -S s Mean(SE)2.(04)2.8(04)§
OO-stCi,
Median41.5 047.80
2.5(08)2.(07) I 1^1.1 1^1 So-
Yes,N(%)4678.0)4(6.7)43(5.)3(59.) No, (%)132.0) (3.)14(2.6)4(0.7) ^'S$d
Question msFTN=59HAN=60LAN=51ACSIN=59 i j 1 J* |*S |I1^Z |§>Z
On the average day, about how many hours do you personaly watch tel vison?
Yes,N{%)30(5.8)27(45.0)23(4.)15(2.4) No, (%)294.)30(5.)30(52.6)4(71.2)
(female)partnershaveyouhadsexwith?'
Now thinking about he time sinceyour18thbirthday,howmanymale < £ Doyouhaveanycreditcardsor charge cards? Please do notQo-S
O include debit cards. 10 cigaret sinyourentirelife?
Have you smoked at least ^

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
930 Lind et al.
Everyda,N(%)294.)23(8.)20(35.1) (8.6) Afewtimsawek,N{%)15(2.4)16(2.7)19(3.)26(4.1) Onceaw k,N{%)4(6.8915.0)8(14.0)6(1.2) Lesthanoceawk,N{%)10(6.9)7(1.)5(8 13(2.0) Nevr,N(%)1 .7 5(83) .8 3(51) Everyda,N(%)71.9) 1(8.3)6(10.5) 1(8.6) Sevraltimesawk,N{%)15(2.4)19(3.7)2(40.)25(4.) Sevraltimesa onth,N{%)16(27.)1(20.)7(12.3) (18.6) Rarely,N(%)1423.7)1(2.7)14(2.6) (10.2) Nevr,N(%)71.9)5(83712.3)6(10.2)
Ap endix B. Continued Question msFTN=59HAN=60LAN=57CASIN=59 How ftendoy ureadthenwspa er? Wouldyoutel mehowoftenyou watchprimetimedram orsitua ion comedy programs?

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Don'tsave,N(%)1830.5)24(0.)19(3.)25(4.) Saveocasinly,N(%)240.7)19(3.7)26(45.)26(4.1) Savergulay,N(%)1728.)17(28.3)1(2.)8(13.6) Nevrin12months,N(%)58. 9(150) 8. 7(19) 1to3imesn12months,N{%)3(5.17 )4(7.058) 4to7imesn12months,N(%)35.1 (83)47.0 (19) 8to1 imesn12months,N(%)58.4(67) 10.5) (8 1to3imesa onth,N{%)13(2.0)13(2.7) 1239(5.) Onceortwicea k,N(%)120.3)7(1 5(26.3)1(5.6) 3to4imesawk,N(%)152.4)1(8.3)14(2.6)5(8 5timesawkorme,N{%)3(5.1467)2(3.50 )
Ap endix B. Continued Question emsFTN=59HAN=60LAN=57 Peoplehavedif er ntreasons forsaving,ev nthoughtheymay notbesavingal thetime.Whic of thefol wingsta ementscomes closest to describing your saving habits?h Thinking back over the last 12months,abouthowregulary di youdrinkalcoh licbev rages?

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
932 Lind et al.
Nota linthepasyer,N(%)712. 4(71)35.47(123) Oncei thepastyer,N(%)23.49(16.)5(89610.5) Atleaswortheimsayer,N(%)132.4)18(32.)18 20(35.1) Onceamonth,N{%)15(2.9)8143 (19.6)2(1.) Onceaw k,N(%)71.98(43)10(7.9)6105 Morethanoceawk,N(%)1423.7)9(16 .1)6(05 Nota linthepasyer,N{%)14(23.7)1(2.0)917 6(27.) Oncei thepastyer,N{%)4(6.810(6.9)(1.3)4(69 Atleas worthe imsayer,N(%)1830.59(2)037.2(46) Onceamonth,N{%)13(2.0)81369(7.0)1(7.2) Onceawk,N(%)58. 4(68)713.2 (7) Morethanoceawk,N(%)58. (5)23.80( ) "Topquartileofresponse,or>=20sexpartnes. bFiveorignaloptinshaveb ncolapsedto hre optins.
Ap endix B. Continued Question msFTN=59HAN=60L =57ACSIN=59 Duringthepast12months,how oftenhaveyouof er dyourseat onabusorinapublicplacetoa stranger who was tanding? Duringthepast12months,how often have you given fo d or money to a homel s person?

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Supplementary Data
Supplementary data are freely available online at http://poq.oxfordjour
org/.
References
Acree, Michael, Maria Ekstrand, Thomas J. Coates, and Ron Stall. 1999. "Mode Effects in Surveys
of Gay Men: A Within-Individual Comparison of Responses by Mail and by Telephone."
Journal of Sex Research 36:67-75.
Aquilino, William S. 1994. "Interview Mode Effects in Surveys of Drug and Alcohol Use: A Field
Experiment." Public Opinion Quarterly 58:210-40.
Aquilino, William S., and Leonard A. LoSciuto. 1990. "Interview Mode Effects in Drug Use
Surveys." Public Opinion Quarterly 54:362-95.
Bateson, Melissa, Daniel Nettle, and Gilbert Roberts. 2006. "Cues of Being Watched Enhance
Cooperation in a Real-World Setting." Biology Letters 2:412-14.
Béland, Yves, and Martin St-Pierre. 2008. "Mode Effects in the Canadian Community Health
Survey: A Comparison of CATI and CAPI." Advances in Telephone Survey Methodology,
edited by James M. Lepkowski, Clyde Tucker, J. Michael Brick, Edith D. De Leeuw, Lilli
Japec, Paul J. Lavrakas, Michael W. Link, and Roberta L. Sangster, 297-314. New York:
Wiley.
Bridges, Janet A. 2000. "A Discussion of Traditional and Computerized Survey Techniques to
Reach Target Publics and a Field Experiment to Determine Their Media Use." Journal of
Promotion Management 5(2): 17-34.
Brittingham, Angela, Roger Tourangeau, and Ward Kay. 1998. "Reports of Smoking in a National
Survey: Self and Proxy Reports in Self- and Interviewer-Administered Questionnaires." Annals
of Epidemiology 8:393—401.
Cassell, Justine. 2000. "Embodied Conversational Interface Agents." Communications of the
ACM 43:70-78.
Cassell, Justine, and Peter Miller. 2008. "Is It Self-Administration if the Computer Gives You
Encouraging Looks?" In Envisioning the Survey Interview of the Future, edited by Frederick
G. Conrad and Michael F. Schober, 161-78. New York: Wiley & Sons.
Cassell, Justine, Joseph Sullivan, Scott Prévost, and Elizabeth F. Churchill, eds. 2000. Embodied
Conversational Agents. Cambridge, MA: MIT Press.
Couper,' Mick P., Roger Tourangeau, and Theresa Marvin. 2009. 'Taking the Audio Out of Audio
CASI." Public Opinion Quarterly 73:281-303.
Ellingsen, Tore, and Magnus Johannesson. 2008. "Anticipated Verbal Feedback Induces Altruistic
Behavior." Evolution and Human Behavior 29:100-105.
Emest-Jones, Max, Daniel Nettle, and Melissa Bateson. 2011. "Effects of Eye Images on Everyday
Cooperative Behavior: A Field Experiment." Evolution and Human Behavior 32:172-78.
Foucault, Brooke. 2009. "Non-Verbal Correlates of Rapport in Face-to-Face Survey Interviews:
An Analysis of Interviewer Behavior." Paper presented at the 2009 Conference of the American
Association for Public Opinion Research, Hollywood, FL, USA.
Fuchs, Marek. 2009. "Gender-of-Interviewer Effects in a Video-Enhanced Web Survey: Results
from a Randomized Field Experiment." Social Psychology 40:37^42.
Groves, Robert M., Floyd J. Fowler Jr., Mick P. Couper, James M. Lepkowski, Eleanor Singer, and
Roger Tourangeau. 2009. Survey Methodology, 2nd ed. New York: Wiley.
Gundersen, Daniel A., Christine D. Delnevo, and Dana C. Momperousse. 2007. "Mode Effects
on Cigarette Smoking Estimates: Comparing CAPI and CATI Responders in the 2001-2002
Current Population Survey." Proceedings of the American Public Health Association Annual
Meeting, Washington, DC, USA.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
934 Lind et al.
Gutierrez-Osuna, Ric
Computer Interact
Berkshire Publishing
Hargittai, Eszter. 20
Computer Review 2
Johnson, Mark H.,
Preferential Trackin
Johnson, Mark H., a
Face Recognition. O
Kanwisher, Nancy, J
A Module in Huma
Neuroscience 17:4302-11.
Kreuter, Frauke, Stanley Presser, and Roger Tourangeau. 2008. "Social Desirability Bias in
CATI, IVR, and Web Surveys: The Effects of Mode and Question Sensitivity." Public Opinion
Quarterly 72:847-65.
Lowe, Richard, and Wolfgang Schnotz, eds. 2007. Learning with Animation: Research
Implications for Design. New York: Cambridge University Press.
Nakamoto, Takamichi, ed. 2013. Human Olfactory Displays and Interfaces: Odor Sensing and
Presentation. Hershey, PA: IGI Global.
Person, Natalie K., Sidney D'Mello, and Andrew Olney. 2008. "Toward Socially Intelligent
Interviewing Systems." In Envisioning the Survey Interview of the Future, edited by Frederick
G. Conrad and Michael F. Schober, 195-214. New York: Wiley & Sons.
Picard, Rosalind W. 1998. Affective Computing. Cambridge, MA: MIT Press.
Preece, Jennifer, Yvonne Rogers, and Helen Sharp. 2011. Interaction Design: Beyond Human
Computer Interaction, 3rd ed. New York: John Wiley & Sons.
Schaeffer, Nora Cate. 2000. "Asking Questions about Sensitive Topics: A Selective Overview."
In The Science of Self-Report: Implications for Research and Practice, edited by Arthur A.
Stone, Christine A. Bachrach, Jared B. Jobe, Howard S. Kurtzman, Virginia S. Cain, and Jaylan
Turkkan, 105-21. New York: Taylor & Francis.
Schober, Michael F., and Susan E. Brennan. 2003. "Processes of Interactive Spoken Discourse: The
Role of the Partner." In Handbook of Discourse Processes, edited by Arthur C. Graesser, Morton
A. Gernsbacher, and Susan R. Goldman, 123-64. Mahwah, NJ: Lawrence Erlbaum Associates.
Sproull, Lee, Mani Subramani, Sara Kiesler, Janet H. Walker, and Keith Waters. 1996. "When the
Interface Is a Face." Human-Computer Interaction 11:97-124.
Steiger, Darby Miller, and Beverly Conroy. 2008. "IVR: Interactive Voice Response." In
International Handbook of Survey Methodology, edited by Edith D. de Leeuw, Joop J. Hox,
and Don A. Dillman, 285-98. New York: Lawrence Erlbaum.
Tourangeau, Roger, Mick P. Couper, and Darby M. Steiger. 2003. "Humanizing Self-Administered
Surveys: Experiments on Social Presence in Web and IVR Surveys." Computers in Human
Behavior 19:1-24.
Tourangeau, Roger, Kenneth Rasinski, Jarred B. Jobe, Tom W. Smith, and William F. Pratt.
1997. "Sources of Error in a Survey of Sexual Behavior." Journal of Official Statistics
13:341-65.
Tourangeau, Roger, Lance J. Rips, and Kenneth Rasinski. 2000. The Psychology of Survey
Response. New York: Cambridge University Press.
Tourangeau, Roger, and Tom W. Smith. 1996. "Asking Sensitive Questions: The Impact of
Data Collection Mode, Question Format, and Question Context." Public Opinion Quarterly
60:275-304.
. 1998. "Collecting Sensitive Information with Different Modes of Data Collection." In

Computer-Assisted Survey Information Collection, edited by Mick P. Couper, Reginald P.
Baker, Jelke Bethlehem, Cynthia Z. F. Clark, Jean Martin, William L. Nicholls II, and James
M. O'Reilly, 431-54. New York: Wiley & Sons.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC
Tourangeau, Roger, and Ting Yan. 2007. "Sensitive Questions in Surveys." Psychological
Bulletin 133:859-83.
Turner, Charles F., Leighton Ku, Susan M. Rogers, Laura D. Lindberg, Joseph H. Pieck, and
Freya L. Sonenstein. 1998. "Adolescent Sexual Behavior, Drug Use, and Violence: Increased
Reporting with Computer Survey Technology." Science 280:867-73.
Tversky, Barbara, Julie Bauer Morrison, and Mireille Betrancourt. 2002. "Animation: Can It
Facilitate?" International Journal of Human-Computer Studies 57:247-62.
von der Pütten, Astrid M., Laura Hoffmann, Jennifer Klatt, and Nicole C. Krämer. 2011. "Quid
Pro Quo? Reciprocal Self-Disclosure and Communicative Accommodation Towards a Virtual
Interviewer." In Proceedings of the 11th International Conference on Intelligent Virtual
Agents, Lecture Notes in Artificial Intelligence 6985, edited by H. Högni Vilhjâlmsson et al.,
183-94. Berlin Heidelberg: Springer.
Yee, Nick, Jeremy N. Bailenson, and Kathryn Rickertsen. 2007. "A Meta-Analysis of the Impact
of the Inclusion and Realism of Human-Like Faces on User Experiences in Interfaces." CHI
2007 Proceedings'. 1-10.

141.211.4.224 on Sun, 03 Jan 2021 11:18:46 UTC

This Content Downloaded From 141.211.4.224 On Sun, 03 Jan 2021 11:18:46 UTC

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

This Content Downloaded From 141.211.4.224 On Sun, 03 Jan 2021 11:18:46 UTC

Uploaded by

Copyright:

Available Formats

WHY DO SURVEY RESPONDENTS DISCLOSE MORE WHEN COMPUTERS ASK THE

Stable URL: https://www.jstor.org/stable/24546177

This content downloaded from

WHY DO SURVEY RESPONDENTS DISCLOSE MORE

Abstract Self-administration of surveys has been shown to inc

doi: 10.1093/poq/nft038 Advance Access publication November 21, 2013

Kesearcn. /\n ngnis reservea. ror permissions, piease e-man: journais.permissionsvyoup.com

This content downloaded from

Survey respondents tend to report more socially undesirable behavior wh

This content downloaded from

Potential Origins of FTF-ACASI Disclosure Differences

This content downloaded from

respondent, the respondent's awareness that the interviewer might be able

1. If sensory technologies (e.g., Gutierrez-Osuna 2004; Person, D'Mello, and Olney 2

This content downloaded from

M 'o _ & S "

This content downloaded from

respondents' willingness to disclose sensitive information is reduced by t

This content downloaded from

This content downloaded from

DESIGN AND PROCEDURE

Respondents were randomly assigned to one of four experimental cond

This content downloaded from

Figure 1. Virtual Interviewer User Interface (with more facial movement

The HA and LA virtual interviewers were created using motion-capture

This content downloaded from

*p C/5 a> tt! C/3 (*> C/2

^ <zj <zj c/5 m <| i/i M m w «3n c« y>

This content downloaded from

Forty-two questions were selected from three well-known and consequential

This content downloaded from

On ten of the questions, respondents in the published studies reported mo

We used ANOVA to test mode differences in the twelve questions r

This content downloaded from

QUESTIONS THAT PRODUCED FTF VERSUS ACASI MODE EFFECTS

Table 4 presents the summary measure (mean, median, or percentage)

TESTING THE THREE HYPOTHESES

We carried out a series of comparisons between our experimental condition

This content downloaded from

High animation CQ Tt <N r-H (SJ pq

This content downloaded from

This content downloaded from

This content downloaded from

This content downloaded from

This content downloaded from

We had complete postinterview questionnaires for 213 of the 235 respon

This content downloaded from

This content downloaded from

How often have you offered your seat on a bus or subway,

Comfortable at start of interview?

Figure 2. Marginal Probability of Giving the Least Socially Desirable

Discussion and Implications

Our findings largely support this proposal. Specifically, they demonstrate

This content downloaded from

This content downloaded from

This content downloaded from

A second attribute useful to know would be the likelihood for a respon

This content downloaded from

— — 'C 32 5/3 "G

sizdef ctsDepnd t3C

This content downloaded from

SurveyofmdeHypothesizdef ctsDepnd tMode f ct

This content downloaded from

p C/5 a> tt! C/3 (> C/2